Practice Development, Artificial Intelligence, Digital Operating Room
Closing the Chasm Between Code and Clinic
The first in a monthly series on AI and ophthalmology.
Bruce Allan MD, FRCS
Published: Tuesday, November 17, 2026
“ It’s a first draft superpower. “
Most of you reading this will be excited by AI or afraid of it, or both. Of the two broad AI categories, generative AI and discriminative AI, generative AI is already well established. The nature of search, for example, has changed thanks to generative AI. Where previously I might have looked up where I could find out how to do a power analysis for a clinical trial, a large language model (LLM) using a suitable prompt can now perform the power analysis for me, explain how to do it, and supply source references in seconds.
Generative AI is all around us, and it is a fabulous tool for evidence summary, preparing patient materials, and consultation summaries. It’s a first draft superpower. But when it comes to making decisions that directly affect patient care, we probably would want to check those source references before we act.
And rightly so. Generative AI models are, after all, simply making a best guess as to the next ‘token’ word or image element to build a response to a prompt. It’s an increasingly good guess, but the risk of failure or ‘hallucination’ is baked in.
Discriminative AI has greater potential for guiding clinical decisions. Discriminative AI bases classification or prediction, the essence of most clinical decision-making tasks, on feature recognition in a training data set. In contrast to the vast banks of unstructured data used to train LLMs, data sets used to train discriminative models are based on labelled data—example data with a description of what it represents.
With enough accurately labelled data, discriminative models can be trained with a tuning algorithm to home in on the features that are clinically important. Because there are no preconceptions about what these features are, AI models can spot important signals, or ‘biomarkers’, that are previously unrecognised, giving discriminative AI models a massive advantage. This is particularly true for so-called deep learning models that are designed to spot patterns in raw data rather than processed tabular data outputs.
Multimodal models combining different types of imaging or combinations of imaging with other biometric data do even better. The net result is that discriminative AI models can match or outperform expert clinicians in many common clinical decision-making tasks. In ophthalmology, applications range from quantifying keratoconus or glaucoma progression risk to screening and classification for retinal disease. Beyond this, the new field of oculomics, in which routine eye imaging is used to screen for general health problems, is making astonishing progress.
You will learn more about all these exciting possibilities in a series of editorials from the ESCRS Digital Health Special Interest Group in upcoming issues. But, if discriminative AI is so good, why aren’t we using it already?
First, we need the labelled data—lots of it. Healthcare data is often siloed in vendor-specific archive systems or is inaccessible because of legal constraints, perceived or real. Interoperable systems are needed to store labelled data in a tractable form, ready for model training, in the cloud. This healthcare data needs to be safely de-identified in most jurisdictions, including Europe, before it can be used for research, while interoperability standards such as DICOM (Digital Imaging and Communications in Medicine), FHIR (Fast Healthcare Interoperability Resources), and SNOMED CT (Systematized Nomenclature of Medicine Clinical Terms) need to be fully developed for ophthalmology for data to flow freely. Navigating these constraints takes time, money, and a lot of effort. Developing high-quality research data sets is doable, but difficult.
The second barrier to wider adoption is generalisability. AI models can only work from the data they are trained on. Stellar performance for a model trained on state-of-the-art imaging data collected in an academic centre may not hold up when the same model is applied in an underserved population screened with older tech. To get around this we need, you guessed it, more training data collected in a wider range of settings. Again, not trivial.
Proving that AI models generalise well with performance evaluation in a range of devices and settings is just one step in the long journey to regulatory approval for new AI as a medical device. The whole product lifecycle must be developed with reference to a suite of quality and safety standards, and the intended use needs clear definition. This requires significant additional funding and know-how. To date, only three autonomous or semi-autonomous ophthalmic AI decision tools have been approved in the US, all designed to automate diabetic retinal health screening.
Despite an equally stringent regulatory environment, around 10 times that number of AI decision tools have approval in Europe, reaching into a wider range of use cases. What started with a trickle will become a flood as the pathways from clinical data curation to modelling to clinic-ready discriminative AI tools become better understood.
Other challenges remain. The more powerful AI models are complex, and their decisions are not easily explainable. But their predictive value can be measured. Based on this, input from everyone with an interest, including target patient groups, is needed to design classification thresholds and prediction tolerances to build public trust in the tools we are using.
Want to understand AI decision tools better yourself? Start by asking your favourite LLM.
Bruce Allan MD, FRCS is a Professor of Ophthalmology and Consultant Ophthalmic Surgeon at Moorfields Eye Hospital, London, UK.